Fairness-Aware Credit Intelligence: Balancing Predictive Accuracy, Inclusion, And Regulatory Accountability

Authors: Dr. Harsha Sammangi, Aditya Jagatha, Ruthwik Gullipalli

Abstract: Machine learning has substantially improved consumer credit-risk prediction, yet its deployment in lending decisions raises persistent concerns regarding demographic fairness, financial exclusion, explainability, and regulatory defensibility. This study develops and empirically evaluates a Fairness-Aware Credit Intelligence (FACI) framework — a four-layer architecture integrating predictive modeling, in- and post-processing fairness intervention, explainability and human policy override, and portfolio-level simulation and governance — using loan-level data from 412,683 consumer lending applications spanning 2021–2026. The study compares a traditional credit scorecard, gradient boosting and deep neural network models, two single-constraint fairness-aware models (demographic parity and equal opportunity), and the integrated FACI framework across predictive accuracy (AUC-ROC), approval rates, demographic parity and equal opportunity differences, disparate impact ratios, portfolio return, and default rates. Results show that unconstrained gradient boosting and neural network models achieve the highest raw accuracy (AUC-ROC 0.781–0.789) but the largest fairness disparities (demographic parity difference 0.187–0.201, disparate impact ratios of 0.65–0.68, below the regulatory four-fifths threshold). The integrated FACI framework achieves AUC-ROC of 0.778 — within 0.003 of the unconstrained gradient boosting benchmark — while reducing the demographic parity difference to 0.038 and improving the disparate impact ratio to 0.92, alongside a higher simulated net portfolio return (5.87%) than either single-constraint fairness model and a lower default rate (8.8%) than the unconstrained benchmark. Subgroup analysis reveals that FACI's gains are concentrated among thin-file applicants, whose approval rate gap relative to the reference group narrows from 27.1 to 14.4 percentage points while their default rate under FACI (10.2%) falls below their default rate under the unconstrained model (11.8%). Portfolio stress simulations across five macroeconomic and operational scenarios demonstrate that FACI's fairness mechanisms function as a form of risk diversification, with smaller return degradation and smaller fairness-metric deterioration than the unconstrained benchmark under severe recession and regional economic shock scenarios. The paper contributes the FACI framework, a five-level maturity roadmap, and a regulatory framework mapping to fintech, responsible AI, and information systems governance research, demonstrating that accuracy-fairness trade-offs documented in prior single-constraint studies can be substantially — though not entirely — resolved through an integrated, multi-layer organizational decision architecture rather than model-level constraints alone.

DOI: http://doi.org/10.5281/zenodo.20842723